RKDOS: A Relative Kernel Density-based Outlier Score

Abstract

This article proposes a novel outlier detection algorithm called Relative Kernel Density-based Outlier Score (RKDOS) to detect local outliers. The proposed algorithm uses a weighted kernel density estimation (WKDE) method with an adaptive kernel width for density estimation at the location of an object based on its extended nearest neighbors. For density estimation, we consider both Reverse Nearest Neighbors (RNN) and k-Nearest Neighbors (kNN) of an object. To achieve smoothness in the measure, the Gaussian kernel function is adopted. Further, to improve discriminating power between normal and abnormal samples, we use an adaptive kernel width concept. Extensive experiments on both synthetic and real data sets have shown that our proposed algorithm has better detection performance over some popular existing outlier detection approaches.